Semi-supervised learning method for heart segmentation model

Abstract

The invention discloses a semi-supervised learning method and system for heart segmentation. According to the method, high-precision heart segmentation can be realized under the condition that only a small amount of labeled data exists through a dynamic pseudo label threshold graph, robust entropy minimization and a contrast consistency strategy. The dynamic pseudo label threshold graph can adaptively select a pseudo label with high confidence, so that the stability and the accuracy of the model are improved. According to the robust entropy minimization strategy, the influence of low-quality labels is reduced by optimizing a cross entropy loss function. And the consistency performance of the model among samples is enhanced by comparing the consistency strategy. Experimental results show that the method is superior to the prior art under different annotation data proportions, and has important clinical application value.

Description

Technical Field

[0001] The present invention relates to the field of cardiac segmentation tasks, and more specifically, to a semi-supervised learning method for a cardiac segmentation model. Background Art

[0002] Currently, the cardiac segmentation task has attracted wide attention and become a key task in medical image analysis. This task aims to accurately extract the cardiac structure from medical images to assist in the diagnosis and treatment of cardiovascular diseases.

[0003] Existing technologies can be divided into three categories: 1) fully supervised learning methods, 2) semi-supervised learning methods, and 3) unsupervised learning methods. The first two methods rely on a large amount of labeled data to achieve high-precision segmentation. However, the annotation process of medical images is usually time-consuming and requires professional knowledge, resulting in difficult and costly data acquisition.

[0004] In recent years, semi-supervised learning methods have begun to receive attention. By using a small amount of labeled data combined with a large amount of unlabeled data, it can alleviate the problem of data scarcity to a certain extent. However, existing semi-supervised methods still face challenges in the confidence evaluation of pseudo-labels and the effective utilization of unlabeled data, which affects the overall performance of the model.

[0005] Unsupervised learning methods attempt to completely get rid of the dependence on labeled data and directly learn from unlabeled data. However, this method often fails to fully capture the subtle features and semantic information of the cardiac structure. Solving these problems will significantly improve the accuracy and practicality of cardiac segmentation. Summary of the Invention

[0006] Aiming at the deficiencies of the existing technologies, the purpose of the present invention is to provide a semi-supervised learning method for a cardiac segmentation model, aiming to use a small amount of labeled data and a large amount of unlabeled data to solve the problem of automatic segmentation of the cardiac structure.

[0007] To achieve the above purpose, the present invention provides the following technical solution: A semi-supervised learning method for a cardiac segmentation model, characterized by including the following steps:

[0008] Step 1, obtain a cardiac image dataset, including a small amount of labeled data and a large amount of unlabeled data;

[0009] Step 2, use a teacher model to generate pseudo-labels, and select high-confidence pseudo-labels through a dynamic pseudo-label threshold map;

[0010] Step 3, train a student model, and optimize it by combining the labeled data and the selected high-confidence pseudo-labels;

[0011] Step 4: Apply the robust entropy minimization strategy to reduce the noise impact caused by low-confidence pseudo-labels by adjusting the cross-entropy loss function;

[0012] Step 5: Adopt the contrastive consistency strategy to ensure that the prediction results of the model are consistent under different augmented views;

[0013] Step 6: Improve the accuracy of pseudo-label generation by iteratively updating the parameters of the teacher model;

[0014] Step 7: Dynamically adjust the threshold strategy during training to adapt to the changes in data distribution;

[0015] Step 8: Finally, combine the supervised loss, robust cross-entropy loss, and contrastive loss to optimize the model parameters and complete the learning of the model.

[0016] As a further improvement of the present invention, the method for obtaining high-confidence pseudo-labels by the dynamic pseudo-label threshold map in Step 2 is as follows:

[0017] Step 2-1: Extract deep features from the unlabeled images, and use the intermediate layer output of the pre-trained teacher model to reflect the local and global information of the images through these features;

[0018] Step 2-2: Use the teacher model to infer the unlabeled data, generate preliminary pseudo-labels, and calculate the confidence score for each pixel;

[0019] Step 2-3: According to the confidence scores of the pseudo-labels, construct a confidence mapping graph, where each pixel value in the graph represents the confidence of the corresponding pseudo-label;

[0020] Step 2-4: Adopt an adaptive algorithm to dynamically adjust the threshold according to the performance of the model during training;

[0021] Step 2-5: Select the pseudo-labels with confidence higher than the threshold by comparing with the dynamic threshold;

[0022] Step 2-6: During training, continuously monitor the model performance, adjust the dynamic threshold according to the feedback, and further optimize the selection of high-confidence pseudo-labels.

[0023] As a further improvement of the present invention, the method for optimizing by combining the labeled data and the selected high-confidence pseudo-labels in Step 3 is as follows:

[0024] Step 3-1: Combine a small amount of labeled data with the high-confidence pseudo-labels selected by the dynamic pseudo-label threshold map to form a training set;

[0025] Step 3-2: Design a comprehensive loss function;

[0026] Step S33: Use the combined training set to iteratively train the student model and optimize the model parameters through backpropagation;

[0027] Step S34: Regularly evaluate the performance of the model on the validation set, monitor the loss value and metrics; and adjust the learning rate and other hyperparameters according to the evaluation results;

[0028] Step S35: During the training process, regularly regenerate the pseudo-labels and update the high-confidence pseudo-labels to ensure that the model adapts to the new data distribution.

[0029] As a further improvement of the present invention, the comprehensive loss function designed in Step S32 includes:

[0030] The labeled data loss function, which uses the standard cross-entropy loss to calculate the prediction error of the labeled data;

[0031] The pseudo-label loss function, which is the loss calculated for the high-confidence pseudo-labels.

[0032] As a further improvement of the present invention, the contrast consistency strategy in Step S5 is specifically as follows:

[0033] Step S51: Apply multiple augmentation techniques to the input image to generate multiple views;

[0034] Step S52: Compare the prediction results of different views and use the consistency loss to minimize the prediction differences of the same instance;

[0035] Step S53: Dynamically adjust the weight of the contrast consistency loss according to the feedback of the training process to further optimize the performance of the model under different views;

[0036] Step S54: Design positive and negative sample pairs during training to improve the discrimination ability of the model by strengthening the similarity between positive samples and reducing the influence of negative samples at the same time.

[0037] As a further improvement of the present invention, after generating multiple views in Step S51, feature extraction is performed on the views, specifically: the model extracts features from different augmented views and calculates the similarity between these features.

[0038] As a further improvement of the present invention, the specific steps for calculating the confidence score of each pixel in Step S22 are as follows:

[0039] Step S221: Introduce a dynamic threshold map M for each class t,c

[0040] M t,c =η c,t ·M t

[0041] Among them, η c,t is the scaling factor for class c, representing the confidence of this class:

[0042]

[0043] Here, ∈ t is the maximum confidence value of different classes;

[0044] Step Two Two Two, the confidence ∈ of the class c,t can be calculated by the following formula:

[0045]

[0046] Among them, Ω c is the mask of class c:

[0047]

[0048] Step Two Two Three, for each pixel (h, w), we calculate its confidence γ(h, w); if the value of class c in t,c is greater than the dynamic threshold M

[0049]

[0050] As a further improvement of the present invention, the formula of the labeled data loss function is as follows:

[0051]

[0052] Among them, is the prediction of the model for the ii-th image. is the corresponding true label. N is the number of labeled data.

[0053] As a further improvement of the present invention, the formula for calculating the loss using high-confidence pseudo-labels is as follows:

[0054]

[0055] Among them, M is the number of high-confidence pseudo-labels, is the high-confidence pseudo-label selected according to the dynamic threshold. As a further improvement of the present invention, the specific steps of applying the robust entropy minimization strategy in Step Four to reduce the noise impact brought by low-confidence pseudo-labels by adjusting the cross-entropy loss function are as follows:

[0056] Step Four One, calculate the robust cross-entropy loss function by the following formula:

[0057]

[0058] Among them, H×W is the height and width of the image, representing the total number of pixels. γ i is the weight for dynamic measurement, reflecting the confidence of pixel i. Pixels with high confidence will contribute more to the loss function. p i is the predicted confidence of pixel ii, indicating the confidence of the model in classifying this pixel;

[0059] Step Four Two, calculate the confidence p through the following formula i :

[0060]

[0061] Among them, is the predicted value output by the model for pixel ii, τ is the temperature hyperparameter, used to adjust the smoothness of the prediction. A higher value will make the prediction distribution more uniform, and C is the number of categories.

[0062] Advantages of the present invention:

[0063] (1) By introducing a dynamic pseudo-label threshold map, we can generate adaptive high-confidence labels for each pixel. This method can not only adjust the label quality in real time but also dynamically optimize according to different stages in the training process, thus significantly improving the accuracy of the model in cardiac structure segmentation. This precise label selection mechanism can effectively reduce the impact of mislabeling and noise on the model performance.

[0064] (2) The framework design of the present invention enables effective training with only a small amount of labeled data. Compared with the large amount of labeled data required by traditional methods, it greatly reduces the cost and time of data collection. This feature is particularly applicable to the field of medical imaging, which has a strong dependence on experts and the annotation process is time-consuming and laborious. Therefore, it is possible to still achieve high-performance model training in the case of scarce labeled data, which has important clinical application value.

[0065] (3) Enhance robustness: By adopting a robust entropy minimization strategy, we can effectively suppress the noise of low-confidence pseudo-labels. This strategy can force the model to focus on high-confidence labels during the training process, thereby enhancing the overall robustness of the model. This method not only improves the adaptability of the model to different datasets but also reduces the overfitting phenomenon when facing new data, ensuring the stability of the model in practical applications.

[0066] In summary, the present invention realizes the generation of adaptive high-confidence labels for each pixel by introducing a dynamic pseudo-label threshold map, greatly improving the accuracy of cardiac structure segmentation. At the same time, the framework design enables effective training with only a small amount of labeled data, significantly reducing the cost and time of data collection, and is particularly suitable for the field of medical imaging. In terms of robustness, the robust entropy minimization strategy effectively suppresses the noise of low-confidence pseudo-labels, enhances the model's adaptability to different datasets, and reduces the overfitting phenomenon. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 is a schematic diagram for adjusting the continuous test time of ASR;

[0068] Figure 2 is a visualization schematic diagram of the dynamic pseudo-label threshold map of an example in the ACDC dataset.

[0069] Figure 3 is a dynamic schematic diagram of the advantages of the three proposed components; namely, the pseudo-label threshold map (DPTM), robust entropy minimization (REM), and contrast consistency (CC). DETAILED DESCRIPTION OF THE INVENTION

[0070] The present invention will be further described in detail below with reference to the embodiments given in the drawings.

[0071] Refer to Figures 1 to 3 As shown, a semi-supervised learning method for a cardiac segmentation model in this embodiment includes the following steps:

[0072] Step 1, obtain a cardiac image dataset, including a small amount of labeled data and a large amount of unlabeled data;

[0073] Step 2, generate pseudo-labels using a teacher model, and select high-confidence pseudo-labels through a dynamic pseudo-label threshold map; Step 3, train a student model, and optimize it by combining the labeled data and the selected high-confidence pseudo-labels; Step 4, apply a robust entropy minimization strategy, and reduce the noise impact brought by low-confidence pseudo-labels by adjusting the cross-entropy loss function;

[0074] Step 5, adopt a contrast consistency strategy to ensure that the prediction results of the model are consistent under different augmented views;

[0075] Step 6, update the teacher model parameters iteratively to improve the accuracy of pseudo-label generation;

[0076] Step 7, dynamically adjust the threshold strategy during training to adapt to changes in data distribution;

[0077] Step 8, finally combine the supervised loss, robust cross-entropy loss, and contrast loss to optimize the model parameters.

[0078] In step 2, a dynamic pseudo-label threshold map is used to select high-confidence pseudo-labels, which can improve the effect of semi-supervised learning. For the sake of convenient description, specifically:

[0079] 2.1) Extract deep features from unlabeled images and utilize the intermediate layer output of the pre-trained teacher model. These features can reflect the local and global information of the images.

[0080] 2.2) Use the teacher model to predict the unlabeled data x t to generate pseudo-labels

[0081] Introduce a dynamic threshold map M t,c

[0082] M t,c = η c,t ·M t

[0083] where η c,t is the scaling factor for class c, representing the confidence of this class:

[0084]

[0085] Here, ∈ t is the maximum confidence value of different classes.

[0086] The confidence ∈ of the class c,t can be calculated by the following formula:

[0087]

[0088] where Ω c is the mask of class c:

[0089]

[0090] For each pixel (h, w), we calculate its confidence γ(h, w). If the value of class c in it is greater than the dynamic threshold M t,c , then this pixel is considered to be of high quality:

[0091]

[0092] 2.3) According to the confidence scores of the pseudo-labels, construct a confidence mapping graph. Each pixel value in this graph represents the confidence of the corresponding pseudo-label.

[0093] 2.4) Adopt an adaptive algorithm to dynamically adjust the threshold according to the performance of the model during training. Statistical methods (such as mean and standard deviation) can be used to set the dynamic threshold to ensure its adaptation to different training stages and data distributions.

[0094] 2.5) By comparing with the dynamic threshold, select the pseudo-labels with confidence higher than the threshold. This step ensures that only high-quality pseudo-labels are retained for training.

[0095] 2.6) During the training process, continuously monitor the model performance, adjust the dynamic threshold according to the feedback, and further optimize the selection of high-confidence pseudo-labels.

[0096] In step 3, when training the student model, we optimize by combining the labeled data and the selected high-confidence pseudo-labels. Specifically as follows.

[0097] 3.1) Labeled data: Combine a small amount of labeled data with the high-confidence pseudo-labels selected through the dynamic pseudo-label threshold map to form a training set.

[0098] 3.2) Design a comprehensive loss function, including: Labeled data loss: Calculate the prediction error of the labeled data using the standard cross-entropy loss. Pseudo-label loss: The loss calculated for the high-confidence pseudo-labels, usually also using the cross-entropy loss, but the weighting coefficient can be adjusted according to the confidence, specifically calculated as the loss through the labeled data. The formula is:

[0099]

[0100] is the prediction of the model for the ii-th image. is the corresponding true label. N is the number of labeled data.

[0101] Use the high-confidence pseudo-labels to calculate the loss:

[0102]

[0103] M is the number of high-confidence pseudo-labels, is the high-confidence pseudo-label selected according to the dynamic threshold. To further improve the robustness of the model, we add a consistency regularization loss:

[0104]

[0105] are the predictions generated for different perturbations of the same input image. Combine the above losses to form the final loss function:

[0106] L = L l + λ 1 L u + λ2 L con

[0107] λ1 and λ2 are hyperparameters used to adjust the influence degree of each part of the loss.

[0108] 3.3) Use the combined training set to iteratively train the student model and optimize the model parameters through backpropagation. During the training process, the model will learn how to better segment the heart structure.

[0109] 3.4) Regularly evaluate the performance of the model on the validation set and monitor the loss value and metrics. According to the evaluation results, adjust the learning rate and other hyperparameters.

[0110] In step 4, a robust entropy minimization strategy is adopted to reduce the noise impact brought by low-confidence pseudo-labels by adjusting the cross-entropy loss function, specifically as follows:

[0111] 4.1) The form of the robust cross-entropy loss function is as follows:

[0112]

[0113] H×W is the height and width of the image, representing the total number of pixels. γ i is the dynamically measured weight, reflecting the confidence of pixel i. Pixels with high confidence will contribute more to the loss function. p i is the predicted confidence of pixel ii, indicating the confidence of the model in classifying this pixel.

[0114] 4.2) The confidence p i is calculated by the following formula:

[0115]

[0116] is the predicted value output by the model for pixel ii, τ is the temperature hyperparameter used to adjust the smoothness of the prediction, and a higher value will make the prediction distribution more uniform. C is the number of classes.

[0117] In step 5, a contrast consistency strategy is adopted to ensure that the prediction results of the model are consistent under different augmented views;

[0118] Furthermore, the contrast consistency strategy is specifically as follows:

[0119] 5.1) Apply multiple augmentation techniques to the input image, such as rotation, scaling, flipping, etc., to generate multiple views. These views are used to train the model to make it robust to changes.

[0120] Feature extraction: The model extracts features from different augmented views and calculates the similarity between these features. By using a contrastive loss function, it ensures that similar inputs are close in the feature space, enhancing the consistency of the model. 5.2) Compare the prediction results of different views and use the consistency loss to minimize the prediction differences of the same instance. This mechanism prompts the model to maintain stable outputs under different input transformations.

[0121] 5.3) Dynamically adjust the weight of the contrastive consistency loss according to the feedback during the training process to further optimize the model's performance under different views.

[0122] 5.4) Design positive and negative sample pairs during training. By strengthening the similarity between positive samples and reducing the influence of negative samples simultaneously, improve the model's discrimination ability.

[0123] In summary, compare the semi-supervised learning method for the cardiac segmentation model in this embodiment with the learning methods in the prior art:

[0124] Comparison 1. The Selftrain model, which is the first proposed unsupervised learning method, uses a basic self-training strategy. The average Dice similarity coefficient (DSC) on the ACDC dataset is 0.713, and on the MMWHS dataset is 0.642.

[0125] Comparison 2. The Dataaug model, which utilizes data augmentation techniques, achieves a DSC of 0.727 on the ACDC dataset and 0.663 on the MMWHS dataset, demonstrating the effectiveness of the augmentation method in model training.

[0126] Comparison 3. The Contextrestoration model, which uses context restoration techniques, obtains a DSC of 0.633 on the ACDC dataset and 0.650 on the MMWHS dataset, indicating its potential in dealing with context information.

[0127] Comparison 4. The Mixmatch model, which adopts an innovative semi-supervised learning strategy, has a DSC of 0.621 on the ACDC dataset and 0.691 on the MMWHS dataset, showing its superiority in utilizing unlabeled data.

[0128] Comparison 5. The Global model, which conducts global feature learning, reaches a DSC of 0.701 on the ACDC dataset and 0.647 on the MMWHS dataset, reflecting the importance of global information for the segmentation results.

[0129] Comparison 6. The Global+Local model, which combines global and local features, has a DSC of 0.757 on the ACDC dataset and 0.710 on the MMWHS dataset, further improving the performance.

[0130] Compared with the SemiContrast model, through the semi-contrast learning method, DSCs of 0.574 and 0.637 were obtained on the ACDC and MMWHS datasets respectively, demonstrating its application potential in an environment with a low amount of labeled data.

[0131] Compared with the PCL model, the DSC is 0.674 on the ACDC dataset and 0.633 on the MMWHS dataset, demonstrating its effectiveness in feature learning.

[0132] Table 1: Comparison of various advanced methods on the ACDC and MMWHS datasets

[0133]

[0134] Compared with the ACINet model, through the co-learning strategy, a DSC of 0.746 was obtained on the ACDC dataset and 0.626 on the MMWHS dataset, showing its advantage in joint task learning.

[0135] Compared with the SSCI model, it is 0.768 and 0.789 on the ACDC and MMWHS datasets respectively, demonstrating the effectiveness of this model in feature extraction.

[0136] Compared with the PatchCL model, the DSC is 0.759 on the ACDC dataset and 0.784 on the MMWHS dataset, demonstrating its ability in local feature modeling.

[0137] For the model (GASLT) of the present invention, when using 10% labeled data, the DSC of the ACDC dataset is 0.904 and that of the MMWHS dataset is 0.735, indicating that it can still achieve performance close to fully supervised learning under low-labeled data conditions, having significant advantages. By using 40% labeled data, GASLT achieved a DSC of 0.856 on the ACDC dataset, further verifying its competitiveness under different data volumes. The method of the present invention performs excellently in the cardiac segmentation task, especially in the case of scarce labeled data, and can effectively improve the segmentation accuracy, demonstrating its broad application potential in related fields.

[0138] In this embodiment, a system carrying the above method is provided, including:

[0139] A data acquisition module for collecting and preparing a small amount of labeled data and a large amount of unlabeled data;

[0140] The teacher model module receives unlabeled image data and adopts a dynamic pseudo-label threshold map strategy. It can dynamically adjust the confidence of pseudo-labels according to data quality, thereby selecting high-quality pseudo-labels. These pseudo-labels are then used as the targets for the student model to assist its learning and optimization.

[0141] The student model module is used for supervised learning to help the model learn real labels. It is used to train by combining labeled data and pseudo-labels and is trained through the pseudo-labels generated by the teacher model to enhance the generalization ability of the model.

[0142] The loss calculation module is responsible for implementing the robust entropy minimization and contrast consistency strategies to optimize the training effect of the model. The loss calculation module can effectively improve the utilization efficiency of pseudo-labels by the model, thereby achieving better performance in semi-supervised learning.

[0143] The iterative update module is used in semi-supervised learning to continuously optimize the model parameters to improve the segmentation performance and stability. It dynamically adjusts the model parameters and threshold strategy.

[0144] The evaluation module is used to verify and test the accuracy and consistency of the model.

[0145] The multi-scale feature extraction module aims to improve the segmentation accuracy of the heart structure. By performing multi-scale convolution processing on the input image, it extracts features at different levels, including detail and global information. The module uses skip connections to fuse low-level and high-level features and applies an attention mechanism to highlight key features and suppress irrelevant information. This feature enhancement method enables the model to better understand the complex heart structure, thereby improving the segmentation effect.

[0146] For the system embodiment, since it basically corresponds to the method embodiment, the relevant parts can be referred to the partial description of the method embodiment. The implementation methods of the remaining modules are not elaborated here. The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of the present invention. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0147] Embodiments of the system of the present invention can be applied to any device with data processing capabilities, and such a device with data processing capabilities can be a device or apparatus such as a computer. The system embodiments can be implemented by software, or by hardware, or by a combination of software and hardware. Taking software implementation as an example, as a logically meaningful device, it is formed by the processor of any device with data processing capabilities where it is located reading the corresponding computer program instructions in the non-volatile memory into the memory for operation.

[0148] The above are only the preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions falling within the concept of the present invention belong to the protection scope of the present invention. It should be noted that for those of ordinary skill in the art in this technical field, several improvements and refinements made without departing from the principle of the present invention should also be regarded as within the protection scope of the present invention.

Claims

1. A semi-supervised learning method for a cardiac segmentation model, characterized in that: The steps include: Step 1: Obtain a cardiac imaging dataset, including a small amount of labeled data and a large amount of unlabeled data; Step 2: Generate pseudo labels using the teacher model and select high-confidence pseudo labels through a dynamic pseudo label threshold map; Step 3: Train the student model and optimize it by combining the labeled data and the selected high-confidence pseudo-labels; Step 4: Apply the robust entropy minimization strategy to reduce the noise impact of low-confidence pseudo-labels by adjusting the cross-entropy loss function; Step 5: Use a comparison consistency strategy to ensure that the prediction results of the model under different enhanced views are consistent; Step 6: Improve the accuracy of pseudo-label generation by iteratively updating the teacher model parameters; Step 7: Dynamically adjust the threshold strategy during training to adapt to changes in data distribution; Step 8: Finally, combine the supervision loss, robust cross entropy loss and contrast loss, optimize the model parameters, and complete the model learning.

2. The semi-supervised learning method for a heart segmentation model according to claim 1, characterized in that: The method for obtaining the high-confidence pseudo-label by using the dynamic pseudo-label threshold map in step 2 is: Step 21: extract deep features from unlabeled images, use the intermediate layer output of the pre-trained teacher model, and use these features to reflect the local and global information of the image; Step 22: Use the teacher model to infer the unlabeled data, generate preliminary pseudo labels, and calculate the confidence score of each pixel; Step 2 and 3: construct a confidence map according to the confidence score of the pseudo label, in which each pixel value represents the confidence of the corresponding pseudo label; Step 24: Adopt an adaptive algorithm to dynamically adjust the threshold according to the performance of the model during training; Step 25: By comparing with the dynamic threshold, select the pseudo label with a confidence level higher than the threshold; Step 26: During the training process, continuously monitor the model performance, adjust the dynamic threshold based on the feedback, and further optimize the selection of high-confidence pseudo-labels.

3. The semi-supervised learning method for a heart segmentation model according to claim 1 or 2, characterized in that: The method for optimizing in step 3 by combining the labeled data and the selected high-confidence pseudo-labels is: Step 31: Combine a small amount of labeled data with high-confidence pseudo-labels selected by a dynamic pseudo-label threshold map to form a training set; Step 32: Design a comprehensive loss function; Step 33, use the combined training set to iteratively train the student model and optimize the model parameters through back propagation; Step 3 and 4: Regularly evaluate the performance of the model on the validation set and monitor the loss value and indicators; And adjust the learning rate and other hyperparameters based on the evaluation results; Step 35: During the training process, pseudo labels are regenerated regularly and high-confidence pseudo labels are updated to ensure that the model adapts to the new data distribution.

4. The semi-supervised learning method for a heart segmentation model according to claim 3, characterized in that: The comprehensive loss function designed in step 32 includes: The labeled data loss function uses the standard cross entropy loss to calculate the prediction error of the labeled data; Pseudo-label loss function, which is the loss calculated for high-confidence pseudo-labels.

5. The semi-supervised learning method for a heart segmentation model according to claim 1 or 2, characterized in that: The specific comparison consistency strategy in step 5 is: Step 51, applying multiple enhancement techniques to the input image to generate multiple views; Step 52: compare the prediction results of different views and use consistency loss to minimize the prediction difference of the same instance; Step 53: According to the feedback from the training process, dynamically adjust the weight of the contrast consistency loss to further optimize the performance of the model under different views; Step 54: Design positive and negative sample pairs during training to improve the model's resolution by strengthening the similarity between positive samples and reducing the impact of negative samples.

6. The semi-supervised learning method for a heart segmentation model according to claim 5, characterized in that: After the step 51 generates multiple views, feature extraction is performed on the views, specifically: the model extracts features from different enhanced views and calculates the similarity between these features.

7. The semi-supervised learning method for a heart segmentation model according to claim 2, characterized in that: The specific steps of calculating the confidence score of each pixel in step 22 are as follows: Step 221: Introduce a dynamic threshold map M for each class t,c M t,c =the c,t ·M t Among them, η c,t is the scaling factor for class c, indicating the confidence of this class: Here, ∈ t is the maximum confidence value of different categories; Step 222, class confidence ∈ c,t It can be calculated by the following formula: Among them, Ω c is the mask of class c: Step 223, for each pixel (h,w), we calculate its confidence γ(h,w); if The value of class c is greater than the dynamic threshold M t,c , the pixel is considered to be of high quality:

8. The semi-supervised learning method for a heart segmentation model according to claim 4, characterized in that: The formula of the labeled data loss function is as follows: in, is the model’s prediction for the ii-th image. is the corresponding true annotation. N is the number of annotated data.

9. The semi-supervised learning method for a heart segmentation model according to claim 4, characterized in that: The formula for calculating the loss using high confidence pseudo labels is as follows: Where M is the number of high-confidence pseudo labels, is a high confidence pseudo label selected based on a dynamic threshold.

10. The semi-supervised learning method for a heart segmentation model according to claim 1 or 2, characterized in that: In step 4, the robust entropy minimization strategy is applied to reduce the noise impact caused by low-confidence pseudo labels by adjusting the cross entropy loss function. The specific steps are as follows: Step 4.1: Calculate the robust cross entropy loss function using the following formula: Among them, H×W is the height and width of the image, representing the total number of pixels. i is the weight of dynamic measurement, which reflects the confidence of pixel i. Pixels with high confidence will contribute more to the loss function. i is the prediction confidence of pixel ii, indicating the confidence of the model in the classification of this pixel; Step 42: Calculate the confidence level p using the following formula: i : in, is the predicted value output by the model for pixel ii, τ is a temperature hyperparameter used to adjust the smoothness of the prediction, a higher value will make the prediction distribution more uniform, and C is the number of categories.

Images